GPU and CPU news and discussions

[Tadasuke]

Intel 4 core 8 thread 2.8 GHz i7-930 in March 2010 was only 6.1% faster than the i7-920 from September 2008. Price was 5 USD higher : $289 instead of $284. The 45nm Bloomfield i7-920 was on the other hand 63.3% faster than the 65nm Kentsfield Q6600 from 2007 ($266 in 2007). +63.3% was much more noticeable than +6.1% as you can imagine.


Then in July, the 45nm Lynnfield i7-875K was only 9.86% faster than the 45nm Bloomfield i7-930 for $342 MSRP, so a whooping 53 USD more (+18.33%), just 4 months later. All reviewers wrote it was too expensive for a four-core CPU and recommended AMD instead.


Radeon HD 6970 came out in December. Was only 11.3% faster than HD 5870 from September 2009, while being 33% more power hungry and 25% more expensive. It was a terrible deal and the first time in the history of graphics cards, when 15 months resulted in such a terribly small improvement (really a decrease in value proposition).


I should had understood then, that Ray Kurzweil and the likes, were wrong. But I didn't. I thought it was only a temporary situation and things would get much, much, much better in the coming years. It took me many years to understand my mistake, my gullibility and my inexperience.

[weatheriscool]

Intel Announces Gaudi 3 AI Accelerator to Compete With Nvidia
Intel says its newest accelerator is superior to Nvidia's current H100 while costing much less.
By Josh Norem April 9, 2024
https://www.extremetech.com/computing/i ... ith-nvidia

In December 2023, Intel held a media event dubbed AI Everywhere. At the end of the event, CEO Pat Gelsinger brandished the company's next-generation AI accelerator, its third-generation Gaudi chip. At the time, Gelsinger said it would be available soon, and that time has come as Intel has made it official at its Intel Vision event today in Arizona.

This is the chip that Intel hopes will allow it to wade into the conversation about AI hardware, as that field has mainly been occupied by Nvidia, which holds between 80% to over 90% of the market, depending on who you ask. AMD also has a very competitive product with its MI300 family of products, and now Intel is entering the fray with Gaudi 3. The company says its newest accelerator delivers up to 50% faster inference than Nvidia's H100 in some workloads and also features 40% better efficiency while costing much less. Oddly, the company's announcement does not mention AMD at all.

[Tadasuke]

Intel is trying to triple total TOPS (tera operations per second) between Meteor Lake (Q1 2024) and Lunar Lake (Q4 2024). AMD also wants to triple TOPS of their NPU in Strix Point, in comparison to Hawk Point, which has 60% more TOPS than AMD Phoenix from Q1 2023. 60% more TOPS translated to 40% higher performance. 3x more TOPS in Intel product can potentially translate to 2x higher performance (than Meteor Lake). Who knows exactly.



source: https://videocardz.com/newz/intel-core- ... ops-for-ai

[Tadasuke]

What truly matters is failure frequency, performance in actual workloads per kilowatt and performance calculated per $1000.

[weatheriscool]

TSMC Reportedly Targeting 2025 and 2027 for Next-Generation 2nm and 1.4nm Processes
Apple will likely be the first customer for both nodes unless something unexpected happens in the next few years.
By Josh Norem April 11, 2024
https://www.extremetech.com/computing/t ... m-and-14nm

Now that TSMC is shifting into high gear with its 3nm process, the discussion turns to the company's next-generation processes, which will be 2nm followed by 1.4nm. Now a new report states TSMC has set its schedule for the highly-anticipated nodes, with 2nm arriving in 2025 and 1.4nm starting to appear somewhere around 2027. Apple will likely be the first customer for both nodes, just like it was with the company's 5nm and 3nm processes.

The new information on TSMC's roadmap comes from Digitimes, so it should be taken with a grain of salt. The article is paywalled, but Macrumors has a summary, stating 2nm trial production will begin later this year, but go into high-volume in 2025. The company will eventually begin 1.4nm production around 2027, but given the timelines involved, that number could change. There was also an earthquake in Taiwan last week that disrupted TSMC's operations for a bit, highlighting how unpredictable life can be on the island. There's also the potential for geopolitical destabilization as well, and TSMC is planning for it by allegedly moving some of its advanced nodes to its US operations.

[Tadasuke]

I made a table comparing memory bandwidth of 17 years of CPU architectures (mostly Intel, but also AMD). As you can see, it does go up exponentially and won't stop anytime soon, but is being improved a bit slowly. It has gone up by 7x since 2008 Penryn, Lunar Lake is going to bring the multiplication to 8x, which is three doublings in 16 years, if LL is available this year.

According to some leaks and rumours, AMD Strix Point (laptop high-end, but possibly also desktop high-end) is supposed to get twice as many memory channels (in the most expensive laptops) than previous AMD laptop CPUs (or standard desktop CPUs) and also twice as many as Intel Lunar Lake, which is allegedly supposed to get the same DRAM speed (DDR5-8533). However, don't expect 273 GB/s on any low-end SKU during this decade. Lunar Lake will get on-package Foveros-integrated 16 or 32 GB of LPDDR5x.

Strix Point will probably have an 8 WGP/16 CUs RDNA 3+ integrated GPU and a next-gen NPU in 64 AI Engine tiles significantly enhancing on-chip AI capabilities. Strix Point APUs could have 24 MB of L3 and 12 MB of L2 cache in total. Its TDP will begin at 28 watts (so won't be for super low-power devices). Memory bandwidth would be useful for both graphical and AI processing. As a point of comparison, Xbox Series S has 4 x 2 GB of 1750 MHz 14 Gbps GDDR6 memory with 128-bit bus, resulting in 224 gigabytes/second (also has 2 GB of 56 GB/s system memory).

My guess is that cheaper 2-channel memory will reach 273 GB/s in Q4 2031 or Q1 2032, seven years after it reaches 273 with 4-channels. It took only 4 years to double bandwidth this time, but it might take longer next time. I find it difficult to believe that we would see 2-channel 240 GB/s in Q1 2028 (but extreme overclocking record is doable). High-end server SKUs already use 12-channel memory (from AMD, 8-channels from Intel), which would result in 819 GB/s total bandwidth with DDR5-8533. DDR6 might come in 2026 or 2027, according to some.

[Tadasuke]

16 lanes of PCI-express 1 allowed for 8 GB/s of bandwidth
16 lanes of PCI-express 5 allows for 128 GB/s of bandwidth
16 lanes of PCI-express 7 is going to allow for 512 GB/s of bandwidth

[Tadasuke]

HBM and AI chips 2022-2025:


Prediction for Q3-Q4 2024:

Current leaks and rumours imply that AMD Zen 5, Intel Arrow Lake, consumer Nvidia Blackwell and new AMD consumer RDNA GPUs will all offer about 25% more performance at the same price as their predecessors (Zen 4, Raptor Lake, Lovelace and RDNA 3) from late 2022.

[Tadasuke]

[Tadasuke]

video about AMD Zen 6 and Zen 7 from "Moore's Law Is Dead" (the guy's name is Tom):

[wjfox]

[weatheriscool]

Samsung to Introduce 3rd Generation Gate-All-Around 2nm Transistors in 2025
The company is already the industry leader when it comes to gate-all-around transistors, but so far it's yet to really make a dent in TSMC's market share.
By Josh Norem April 30, 2024

Samsung was the first fab to launch a 3nm process in mid-2022, beating TSMC to market by about six months. Plus, its 3nm node offers gate-all-around (GAA) transistors, which none of its rivals have launched yet. Despite these advantages, Apple chose TSMC for its M3 SoCs. Samsung has yet to announce any major contracts for its 3nm GAA chips, even though they're almost two years old now. Nonetheless, Samsung is pressing forward with GAA transistors and will launch the third generation of this technology with its 2nm process in 2025.

News of Samsung's plans comes from Business Korea, which has a new report on Samsung's efforts to compete with rivals TSMC and Intel. It states the obvious upfront: Samsung is the only fab with commercialized gate-all-around transistors, but that effort has failed to move the needle. It lays the blame for this situation on the global economic downturn in 2022 and 2023, high production costs, and a limited number of clients interested in bleeding-edge technology. One big factor it fails to mention is Samsung's struggles with yields, which likely caused top-tier clients like Apple and Nvidia to shop at TSMC instead.

https://www.extremetech.com/computing/s ... sistors-in

[Tadasuke]

At least for those living in the UK, you are now able to buy online Radeon 7900 XTX for 800 £ while Radeon 7800 XT is available for 490 £.

On August 20th 2021, Asus Radeon RX 6900 XT TUF Gaming OC was selling for 1350 £, which was quite a lot, when taking into consideration that back in March of 2017, GeForce GTX 1080 Ti was selling (on launch) for only 671 £, which is about half of that price, probably about half performance and 69% of available memory. So between 2017 and 2021, performance/price and memory/price has not improved at all in graphics cards.

Here's some UK price history of ASRock Radeon RX 7900 XTX Phantom Gaming OC for a comparison:


At least in 2024, GPUs finally have some clear increase in value proposition, compared to April of 2017, when GPUs were the cheapest before 2019. Back then:
• new GeForce GTX 1050 Ti 4GB was selling for $140
• new Radeon RX 570 4GB was selling for $170
• new Radeon RX 580 8GB was selling for $230
• new GTX 1060 6GB was selling for $250
• new GTX 1070 8GB was selling for $360
• new GTX 1080 8GB was selling for $500
• new GTX 1080 Ti 11GB was selling for $700
• new Titan XP 12GB was selling for $1200 and it was the top-tier card

[weatheriscool]

TMSC Announces Low-Power N4e Node, Says It's Boosting Specialty Nodes by 50%
The company is redoubling its efforts on nodes other than its main moneymakers.
https://www.extremetech.com/computing/t ... y-nodes-by

TSMC held its European Technology Symposium this week, which you'll be shocked to hear was titled, "Powering AI with Silicon Leadership." The company does a kind of roadshow from April through June each year, hosting symposiums around the world to show its wares to customers. At the European round this week, the company revealed plans for a new low-power node called N4e, which was not on its roadmap previously. It also stated it will boost its specialty technology fab capacity by 50% in the coming years, showing TSMC is focusing on a lot more than just its most advanced nodes.

[Tadasuke]

EPYC 4004 series are 4 to 16 cores CPUs for the AM5 platform, TDP 65 to 170 watts, clocked anywhere from 3.7 to 5.7 GHz, prices anywhere from 149 to 699 USD. As we can see, they seem to be significantly above current low-end Intel Xeon in performance, efficiency and value proposition.

[Tadasuke]

Average Time Spy score of GPUs which people test has risen from about 5,000 in Q1 2016 to to about 14,000 in Q3 2023, that's 2.8x growth in 7 and a half years.

[weatheriscool]

Samsung Is Moving 1nm Production Up One Year to 2026: Report
The company is expected to announce its ambitious plans in June.
By Josh Norem May 30, 2024

Just as CPU companies raced to reach 1GHz in 1999, semiconductor foundries are competing to be the first to produce a 1nm transistor. TSMC isn't expected to hit that high watermark until 2030 or so, and Intel is likely a few years ahead of that. But now Samsung reportedly plans to leapfrog them both by announcing that 1nm production will start in 2026, one year ahead of schedule. However, this is 1.4nm, not precisely 1nm, but it would make Samsung the first fab to go smaller than 2nm.

https://www.extremetech.com/computing/s ... 026-report

[weatheriscool]

[wjfox]

[weatheriscool]

Intel Reveals Gaudi 3 Pricing, and It's Roughly Half the Cost of Nvidia Blackwell, Hopper
The company is hoping aggressive pricing will allow it to get a beachhead in the AI arms race.
By Josh Norem June 7, 2024
https://www.extremetech.com/computing/i ... -blackwell

This week, Intel revealed pricing for its Gaudi 2 and 3 AI accelerators, which are extremely competitive compared to Nvidia's offerings. At the Computex trade show in Taipei, Intel CEO Pat Gelsinger unveiled a pricing sheet, a rare event in the AI world as pricing isn't typically disclosed to the public. Gelsinger revealed that an individual Gaudi 3 accelerator costs around $15,000, roughly half the cost of competing GPUs from Nvidia.

According to slides posted on X by analyst Jim McGregor, a Gaudi 2 accelerator kit costs $65,000, including eight accelerators on a baseboard. The same kit in Gaudi 3 trim, again with eight cards mounted to a baseboard, will go for $125,000, so an individual Gaudi 3 card costs roughly $15,625. For context, that's approximately half the cost of an older Hopper H100 GPU from Nvidia, though it's hard to know exact pricing since they're sold in bulk and usually through private commercial deals. Intel confirmed these prices as well.